| In recent years,multi-label learning has been used more and more in the fields of text classification,protein function prediction,and image annotation.Among the various multi-label applications,the most important thing is to apply the corresponding labels which are correctly classified.Like traditional machine learning,multi-label learning also faces the problem of dimensional disaster,so multi-label feature selection comes into being,however,unlike traditional machine learning,in multi-label machine learning,there may be certain correlations and differences between labels,and there may be certain correlations and redundancy between features.Although most of the existing multi-label feature selection methods take into account the correlation between the labels,however,it does not take into account the correlation between features in the feature space,or considering the correlation in the feature space,but fail to effectively remove redundant features in the feature space,at the same time,the difference of labels is also ignored.In view of the above problems,this thesis proposes a Multi-Label Feature Selection Method based on Feature Group Correlation Analysis(MLFS-FGC)to remove redundant features in the feature space.At the same time,a Multi-Label Feature Selection Method based on Feature Space Fusion(MFS-SF)is proposed,which fuses the shared feature space and the label-specific feature space to ensure that the individuality and relevance between labels are considered.The main research work of this thesis is as follows:1.In order to better mine the correlation and redundancy between features and features in the sample feature space,MLFS-FGC is proposed.This method first groups the features in the feature space into groups;use the information gain and information entropy to calculate the correlation between features in the feature space,if it is greater than a certain value,there is a strong correlation and they are divided into the same group;otherwise,they are not in the same group,after the grouping is completed,there is a strong correlation between features within the same group,and a weak correlation between features between groups.Then use Laplace score to score each feature in the grouped feature group.Finally,for all feature groups,top-k features are extracted according to the size of the feature score,and a batch of features with low correlation are filtered out.In this way,the redundant features in the shared feature space can be removed,and the purpose of improving the classification effect and reducing the feature dimension can be achieved.2.In order to further explore the problem of label differences in the sample label space,MFS-SF is proposed.This method first performs a positive-negative clustering is performed on the sample according to the label attribute corresponding to the sample.Then,the number of clusters is determined separately for the positive and negative clusters composed of positive and negative samples,and the distance between the original feature and the center of each class in the positive and negative clusters is calculated,in this way,a specific feature space of the label is generated.Finally,the shared feature space and the label-specific feature space are fused to ensure that the individuality and relevance between multiple labels are considered,and the problem of label differences is solved.Experiments on 9 public multi-label data sets show that compared with the existing multi-label feature selection methods,the method in this thesis has better performance on each classification index.After adding feature space fusion,on each classification index,it has been further improved,which fully proves the effectiveness of this method. |
